Image restoration via N-nearest neighbour classification

Abstract

A novel and powerful perspective on image reconstruction and restoration is to regard the computational objective as the classification of corrupt (=unclassified) pixels using the classification of the nearest uncorrupt (=classified) pixels. In N-nearest neighbour (NNN) restoration, the distance transform is used to determine the set of N-or-more classified pixels which are as close, or closer, than the Nth nearest to each corrupt pixel. NN classification includes classic restoration algorithms, but new algorithms are implied, especially for colour and gray-scale images that are very sparse or highly corrupt. We present experimental results for an NNN restoration algorithm, for N=1, using for nearest set classification the median of the one-or-more nearest 'good' neighbours. At low corruption levels this algorithm is equivalent to classic median filtering; for images with random pixel loss of 50% to 90%, satisfactory restoration has been achieved for both gray-scale and colour images.